1. CIS 391 - Intro to AI 1 Interpreting Line Drawings (An Introduction to Constraint Satisfaction) Mitch Marcus CIS 391 Fall 2008

2. CIS 391 - Intro to AI 2 We Interpret Line Drawings As 3D  We have strong intuitions about line drawings of simple geometric figures:  We naturally interpret 2D line drawings as planar representations of 3D objects.  We interpret each line as being either a convex, concave or occluding edge in the actual object.

3. CIS 391 - Intro to AI 3 Possible Vs. Impossible Line Drawings  We can reject some line drawings as impossible 3D objects, but only after some confusion. “The Devil’s Trident”

4. CIS 391 - Intro to AI 4  We will consider scenes made up only of opaque trihedral objects, objects where All object faces are planar Exactly three faces intersect at each vertex Interpreting Trihedral Scenes

5. CIS 391 - Intro to AI 5 Convexity Labeling Conventions Each edge in an image can be interpreted to be either a convex edge, a concave edge or an occluding edge:  + labels a convex edge (angled toward the viewer);  - labels a concave edge (angled away from the viewer);  labels an occluding edge. To its right is the body for which the arrow line provides an edge. On its left is space. convexconcave occluding

6. CIS 391 - Intro to AI 6 Huffman/Clowes Junction Labels  A trihedral image can be automatically interpreted given information about each junction of three lines in the image.  Each interpretation gives convexity information for each junction,  This interpretation is based on the junction type. arrow Y L T junction junction junction junction

7. CIS 391 - Intro to AI 7 Arrow Junctions have 3 interpretations (from Winston, Intro to Artificial Intelligence) A1 A3 A2 Important: The image of a given vertex gives a different junction type from a different point of view

8. CIS 391 - Intro to AI 8 L Junctions have 6 interpretations L3 L1 L6L5 L2 L4

9. CIS 391 - Intro to AI 9 The Same “Arrow” and “L” Junctions in a Different Orientation

10. CIS 391 - Intro to AI 10 “T” Junctions Have 4 Interpretations, “Y” Junctions Have 5 + -

11. CIS 391 - Intro to AI 11 Summary Type of JunctionNumber of Physically Possible Interpretations Arrow Junctions3 L Junctions6 T Junctions4 Y Junctions5

12. CIS 391 - Intro to AI 12 Huffman/Clowes Line Drawing Interpretation  Given: a line drawing of a simple “blocks world” physical image  Compute: a set of junction labels that yields a consistent physical interpretation  Assuming: the General Viewpoint Assumption

13. CIS 391 - Intro to AI 13 The General-Viewpoint Assumption:  Shifts in the position of the viewer do not affect the configuration of the line drawing.  This rules out the possibility of the accidental alignment of image features into a spurious junction. (drawings from Machine Interpretation of Line Drawings, Kokichi Sugihara, MIT Press)

14. CIS 391 - Intro to AI 14 The Edge Consistency Constraint

15. CIS 391 - Intro to AI 15 The Edge Consistency Constraint Any consistent assignment of labels to the junctions in a picture must assign the same line label to any given line. L4 + + L3... + - L3... L5

16. CIS 391 - Intro to AI 16  Consider an arrow junction with an L junction to the right:  A1 and either L1 or L6 are compatible since they both associate the same kind of arrow with the line.  A1 and L2 are incompatible, since the arrows are pointed in the opposing directions,  Similarly, A1 and L3 are incompatible. An Example of Edge Consistency

17. CIS 391 - Intro to AI 17 An Example of Edge Consistency II  A2 is compatible with ___?  A3 is compatible with ___? Ans: (A2, L4) (A3, L5)

18. CIS 391 - Intro to AI 18 The Generate and Test Algorithm 1.Generate all conceivable labelings. 2.Test each labeling to see whether it violates the edge consistency constraint; throw out those that do. Complexity:  Each junction has on average 4.5 labeling interpretations.  If a figure has N junctions, we would have to generate and test 4.5 N different labelings.  Thus the Generate and Test Algorithm is O(4.5 N ).

19. CIS 391 - Intro to AI 19 Generate & Test: A Useful Algorithm??  A picture with 27 junctions, like the devil’s trident) will not be rejected until all 4.5 27 hypotheses have been rejected, leaving no interpretation.  4.5 27 = 433249302231073824.244664378464222  A computer capable of checking for edge consistency at a rate of 1 hypothesis per nanosecond would take about 1,000 years to establish that the devil’s trident has no consistent interpretation!

20. CIS 391 - Intro to AI 20 Improving on Generate & Test  The Generate and Test algorithm considers too many hypotheses that are simply outright impossible according to the edge consistency constraint.  The Solution:

21. CIS 391 - Intro to AI 21 Locally Consistent Search  A better algorithm would exploit locally consistent labelings to construct a globally consistent interpretation: 1.No junction label would be added unless it was consistent with its immediate neighbors; 2.If the interpretation could not be continued because no consistent junction label can be added, that interpretation would be abandoned immediately.

22. CIS 391 - Intro to AI 22 Search Trees to the Rescue! Implementation by search tree: 1.Select some junction as the root. 2.Label children at the 1 st level with all possible interpretations of that junction. 3.Label their children with possible consistent interpretations of some junction adjacent to that junction. 4.Each level of the tree adds one more labeled node to the growing interpretation. 5.Leaves represent either futile interpretations that cannot be continued or full interpretations of the line drawing.

23. CIS 391 - Intro to AI 23 The Top Three Levels of a Search Tree A1A3A2A1 L1 A1 L5 A2 L3 A3 L6

24. CIS 391 - Intro to AI 24 The Fourth Level of the Search Tree A1A3 A2 A1 L1 A1 L5 A2 L3 A3 L6 A1 L1 A1 L5 A3 A2 L3 XX A3 L6 A1 X

25. CIS 391 - Intro to AI 25 Another Idea… We are doing somewhat better than the Generate and Test algorithm by not examining useless paths. Nevertheless, we are not fully exploiting the structure of the problem:  Consider a picture with disconnected “islands” (junctions which are not connected by a path along any series of edges).  All our current search methods will explore and re-explore these islands even though they do not interact with the rest of the picture!! This is very wasteful…..